In this section we provide the reader with a comprehensive definition of price discrimination by describing the economic theories and by illustrating several state-of-the-art previous studies. Additionally, we introduce the airline tickets market by explaining its peculiarities.
2.1. Price Discrimination
According to J. Tirole [1
], price discrimination occurs whenever the same commodity is sold at different prices, either to the same customer or to different customers. More precisely, G. Stigler [2
] states that price discrimination arises when two or more similar goods are sold at prices that are in different ratios to marginal costs. The latter definition rules out the differences in prices determined by the differences in costs of serving different customers. Moreover, it includes the case in which the seller discriminates by setting a uniform price.
The conventional economic theory adopts the classification made by [3
] and distinguishes among three different categories of price discrimination. Adopting first-degree price discrimination
(a.k.a. personalized pricing
), the seller is able to set a price equal to the willingness-to-pay (WTP) of each consumer so to extrapolate the entire social surplus. This strategy has often been considered as a useful but abstract benchmark. However, big data and the behavior-based price discrimination are rewriting this wisdom: as highlighted in [4
], first-degree price discrimination is nowadays a more realistic framework since each single user produces enough personal data for the seller to infer about one’s tastes and possibilities. Similarly, third-degree price discrimination
) occurs when different prices are assigned to different groups of customers. Each group is composed by customers who share some common features (e.g., business vs. leisure travelers). In other words, this kind of discrimination is a less granular form of first-degree price discrimination: that is why the more modern categorization addresses both of them as direct price discrimination
. Finally, the remaining category is the second-degree price discrimination
), that is the indirect price discrimination
, and it is based on indirect signals. In this case, the seller lacks of information, and so sets a menu of prices and offers among which the consumer can choose. By properly designing these bundles, consumers will have an incentive to truthfully reveal their preferences and WTPs (see [5
]). The offers may differ either in quantity or in quality. In line with [6
], several kind of indirect price discrimination strategies may be identified: coupons, quantity discounts, bundling, performance-based discrimination, restrictions on purchase and use, knowledge-based discrimination, and non-linear pricing.
It is worth noting that the advent of data analysis and the development of behavioral marketing have triggered the rise of a new category of price discrimination strategies: the behavior-based price discrimination (BBPD)
. According to [7
]: sellers are now using big data and digital technology to explore consumer demand, to steer consumers towards particular products, to create targeted advertising and marketing offers, and in a more limited and experimental fashion, to set personalised prices. At the same time, buyers are making use of the Internet and the variety of choices and tools it provides to ensure that they get a good deal
. For sure, BBPD can be defined as a more sophisticated kind of direct and interpersonal price discrimination. However, it can also be seen as a mix of strategies whose aim is to indirectly incentivize the consumer to autonomously select her type. Economic scholars have widely scrutinized these practices. Among them, add-on pricing scheme
and price obfuscation strategies (see [8
]), price inertia (see [9
]), and price dispersion (see [10
]) are widely adopted in online flight booking too. All these strategies can be traced back to the behavioral biases that affects human behavior, like information overload, status quo bias, loss aversion, framing, hyperbolic discounting. A good summary of the ones which behavioral marketing usually refers to is included in [11
Before going on, it is important to distinguish personalized from dynamic pricing. The only common feature is the use of new technologies and the goal, i.e., to maximize firm’s profit. This net distinction has been very well studied in all its facets by [12
]. Dynamic pricing is that practice through which online seller use algorithms to constantly update their prices in line with the information related to both the demand and the supply side of a market. Since the information that sellers can get online are abundant and it is becoming increasingly cheaper, then it is quite common to elaborate these data to infer on the market conditions in order to choose a price that can more likely maximize producers’ welfare. Anyway, dynamic pricing does not involve any kind of discrimination, since it is not related to a specific target of consumers but concerns the whole market and its equilibrium.
This distinction is crucial when one studies a market such as the online flight booking one. Indeed, ticket prices are constantly updated according to many factors such as seat availability, season, country of purchase, length of stay, and so on. These variables make ticket prices quite volatile and dynamic.
There are three major underlying conditions whose absence would prevent the seller from personalizing prices [13
]. First of all, the seller must have the opportunity to set a price, i.e., she must have some market power. If no market power is in the hand of the seller, then there is no chance to make any decision about the price of any product. Second, a no arbitrage
condition must holds. Indeed, it is necessary to prevent buyers from generating secondary markets in which the customer whose WTP is low resells the commodity to one whit a higher WTP. Indeed, under price discrimination the low type—i.e., the buyer with a low WTP—pays a lower price than the one paid by a high type—i.e., the buyer with a high WTP- and this opens to secondary markets. If this transferability of the commodity
is feasible, then all the effort of the seller would be useless and this strategy would not be profitable anymore. Third, it has to be possible to obtain information about the customers and, more in general, about the structure of the demand side. If not, it would be nearly impossible both to evaluate differences among consumers and to segment the market.
For what concerns the online airline market, it is reasonable to expect some price discrimination to arise. As it will be better explained later, this market is an oligopolistic one, where firms have some market power. Moreover, they are able to collect information about the demand side. Filling this gap has been considered an important part of this business since the birth of the airline market, and the ability in gathering information has sharply increased with the advent of the digital economy. Furthermore, flight tickets are very difficult to resell because of their specificities, and the increasing personalization makes this practice even harder.
Looking at the possible effects, in general price discrimination is perceived as a welfare-enhancing strategy, and so regulatory agencies tend not to ex-ante prevent it.
As explained in [14
], a major role is played by the market expansion effect
. By price discriminating, the producer increases the amount of output, and sets up a price schedule that makes that good affordable also for previously excluded customers. As a second positive side, personalized prices lower the search cost that the customers should normally face (see [10
]). Moreover, if one considers a non-monopolistic setting, then price discrimination strategies may strengthen competition. As mentioned in [15
], the possibility to address customers with targeted discounts may reduce the barriers to entry. At the same time, price discrimination can be read as a segmentation of the market that generates a conspicuous amount of smaller and less heterogeneous niches of customers, and it is possible that the original barriers to entry are not strong enough to prevent the entrance of a new competitor in a single fraction of the original market.
However, as described in [16
], there might be cases in which price discrimination may harm social welfare and especially consumer surplus. These cases are much more likely to arise in (quasi-)monopolistic markets [15
]. The most evident negative side is the so-called appropriation effect
: contrary to the market expansion effect, this is that exploitative practice through which the seller offers a price higher than the uniform one to those already-in-the-market customers whose WTP is high enough (e.g., loyal customers). In this way, the seller is able either to counterbalance the rise in costs due to the increase in the size of the output or simply to extract some rent from the clients. Moreover, price discrimination may be harmful when it is not transparent. In this case, consumers’ behavioral biases may be leveraged to artificially shift up the naive customers’ demand. At the same time, the lack of transparency may incentive the breach of the rules of fairness, such as the prohibition of discrimination based on gender, race, or disability (see, for example, [17
]). In general, the use of this practice may reduce the level of trust inside the market. Finally, considering a non-monopolistic market, personalized pricing may be harmful whenever it is adopted by a super-dominant firm as an exclusionary practice to force competitors out of the market (e.g., predatory pricing, fidelity rebates, loyalty discounts, bundled discounts and rebates, margin squeeze). In this situation, the cost to discriminate may be too high for the competitors and the beneficial effect may be too low to balance the loss of competition.
2.2. Previous Studies
When analysing the economic literature on airline ticket pricing, the first important result to highlight is the huge volatility. The impact of several dynamic factors—flight distance, fuel price, air traffic, flight classes, seasonality—generate fares which differ even for seats on the same flight. Therefore, a first strand is devoted to the study of dynamic pricing. However, addressing this phenomenon as discriminatory may be a mistake. Indeed, in most cases the change in price is simply the result of an update of the available market-related information. Nevertheless, evidence has been found about what has been defined as dynamic price discrimination
]. According to the previous definitions, it can be seen as a mix between grouping and product differentiation. The typical example is the trend of the airline companies to offer higher prices during office hours and lower prices in the evening. This scheme is a screening mechanism through which the consumers have incentive in autonomously reveal their types: business travelers tend to buy while they are at work and they have a low elasticity of demand, while leisure travelers usually buy a ticket in the evening and they are characterized by a more elastic demand. As a consequence, economic scholars have focused their attention on the perfect timing theory to buy a plane ticket. Among the others, the authors of [20
] analyse Russian airline ticket market and compares local and global flights price behavior for the spring-summer 2015 period in the two main hubs in Russia (i.e., Moscow and Saint Petersburg) selecting 50 most popular destinations from Moscow. For each day a request to get the minimum price has been done. In line with the results obtained in previous works, they conclude that it is better to buy either in advance to prevent price increases in the future or few days before departure. However, this result is not valid for internal flights, which are highly influenced by the lack of competition and the absence of low-cost carriers. Similarly, the authors of [21
] testifies how the lack of information on the companies’ fares makes difficult for the buyer to determine the perfect purchase time, even when historical data series are available. By collecting consumers-available data and using a lag scheme to include lagged features, the authors build a PLS regression model to predict prices. The results obtained through the experiment show that buying as early as possible is not always the best policy. For example, airlines can change fares until the last moment, lowering prices either to increase sales or to fill unsold seats. Therefore, committing to a specific ticket a long time before the flight may be risky, and the use of consumer algorithms may be cost-saving and beneficial. In line with this reasoning, the authors of [22
] built a forecasting system to help consumers in purchasing tickets by combining an ARMA algorithm and a random forest algorithm. By using data from nine cities in China and taking into account crucial variables such as take-off times, departure date, and competition from other airlines, they demonstrate how this model can be effective in predicting future prices. Another interesting forecasting model has been developed in [23
] by means of Machine Learning techniques. Once collected data about twenty flights between the 5 major American hubs (Atlanta, Chicago, Los Angeles, Dallas, Denver) for one week, they run a model to predict the price at a future date, the minimum value of a fare, and the expected fluctuations of the price. In conclusion, it is necessary to remind that the big flaw of these systems is the lack of data: even if the information of the average ticket price can be extracted from travel sites, the one regarding the prices of specific flights or the number of available seats on the flight are not always made public either for reasons of competition or because of private negotiations.
Apart from this first block of research, a second part of the literature focuses on the search for other and more classic forms of price discrimination in the digital markets. Our paper enters this debate. The leading works about online price discrimination are [24
] since they found the first empirical results by using pioneering empirical methods. Once collected the data, these researchers try to measure price discrimination by ruling out any possible source of noise. Indeed, two prices assigned to the same product may diverge for reasons other than the difference in the willingness to pay (WTP): technical factors such as the distributed infrastructure or the update of the search index may be a first explanation, as well as A/B testing and the differences in geolocation that, in turn, determines differences in monetary conversion, taxation, and shipping costs. Looking at the market we are studying, an interesting phenomenon is the caching of prices
: for cost reasons indeed user data is stored in the cache, which records the essential information of a site to use it faster if needed. This can cause a spread between the price that buyer sees the first time on the site and final price when purchasing the ticket. Moreover, one should also be able to disentangle personalized prices from dynamic price discrimination. Once the noises are silenced, any remaining difference has to be attributed to a different WTP. However, in most cases price discrimination is absent. Apart from the famous case of Amazon’s DVDs in 2000 and interesting journalistic investigations ([28
]), several studies have demonstrated that online price discrimination exists even if it is rare and hard to measure (see [30
]). Among the others, in 2011 TheTrainline.com, Expedia, Easyjet, Virgin, Lastminute and Eurostar were accused of price discrimination based on previous research or queries. By contrast, strong empirical evidence about search discrimination
have been found. For example, Ref. [28
] reported that Orbiz, an online travel agency (OTA) first offered Apple laptop owners the most expensive hotel and travel results. Similar works have focused the attention to the online airline ticket market, since the advent of the new technologies has had an important impact on this market (e.g., [33
]), as it will be shown in the subsequent paragraph. Particularly, Ref. [34
] is the one that more closely resembles the approach we adopt. It emulated the practice of searching for users for a specific flight. In doing so, they run a three-week experiment, analysing 25 companies with a dozen profiles (with three different user’s profiles: affluent, budget and flight comparer), deactivating or activating the tracking systems (cookies) in different browsers (Chrome, Safari, Internet Explorer) and in different geolocations (using two IPs), and they produced 130,000 queries. This type of research was carried out twice a day for each Airlines in a consecutive and non-simultaneous fashion so to avoid blocking the airline’s website server. In line with the more general results, no evidence was found of price discrimination phenomena between the different profiles.
2.3. Airline Tickets Market
Before introducing the empirical analysis, it is important to briefly summarize the structure of the airline tickets markets to fully understand the subsequent steps. As shown in Figure 1
, apart from travelers (i.e., the demand side), the airline tickets market mainly consists of: airline companies (e.g., Alitalia, Rome, Italy; Lufthansa, Cologne, Germany), Global Distribution Systems (GDSs) (e.g., Amadeus, Madrid, Spain; Sabre, Southlake, TX, USA; Travelport, Langley, Slough, United Kingdom), online travel agencies (OTAs) (e.g., Viator, San Francisco, CA, USA; Expedia, Redmond, WA, USA), content aggregators (e.g., Skyscanner, Edinburgh, Scotland, UK; Kayak, Stamford, CT, USA), and consolidators (e.g., Mondee Group, Silicon Valley, CA, USA). They tend to adopt common standards and protocols to share information and regulate the market. Traditionally, the airline companies use to set prices and tariffs according to the flights information, and then to transfer these offers to the Global Distribution System (GDS). These few players, in turn, use their network to transmit these offers to the travel agents. Among the remaining players, the GDSs usually interact with OTAs, which offer customers the possibility to comprehensively organize their travel, from flights booking to hotels. Differently than OTAs, the content aggregators are search engines that help customers in comparing all the available offers and then redirect them to the service provider. Finally, the consolidators are middlemen that sign private agreements on tickets tariffs with the airline companies, and resell them to large companies, or OTAs. Figure 1
provides a schematic overview of the ticket distribution channels.
However, the advent of big data and AI is remarkably reshaping this market. Because of its characteristics, the aviation sector was one of the first to make investments in analytical and intelligent technologies that include data analysis, dynamic dashboards, classification and targeting techniques to increase efficiency in airport management and to optimize costs. The empirical analysis conducted by the authors of [35
] states that airlines need to adopt a data mining systems due to the huge amount of data generated. This study proves how the data infrastructure and series analysis help airlines in sales and logistics management. Indeed, airlines that adopt data analytics are able to offer better sale conditions to consumers, to change the prices dynamically and to avoid unsold seats. The second aspect concerns marketing: data mining allows to customize customers’ travel experience, manage delays and reschedule reservations. In other words, airlines have the possibility to both differentiate their products and better personalize customers’ experience and offers by including discounts and ancillaries. This returns positive effects to the companies in terms of brand loyalty, customer satisfaction and increase in revenues. The third aspect is related to the improvement of airport management: data analysis tools helps to analyse passenger flows, the time spent between arrival and take-off. It makes possible to maintain a high-quality standard, which is also useful for increasing safety systems. The fourth aspect concerns the efficiency in terms of costs: intelligence tools enable companies to analyse the atmospheric conditions or flight speed, reducing times and fuel costs. Finally, having access to historical data improves the booking process and the proposal of accessory services (favorite places, hotels, and transportation). In order to fully exploit the use of big data and machine learning (ML), in 2012 the International Air Transport Association (IATA) has replaced the old EDIFACT protocol with a new API standard, the New Distributed Capability (NDC). The aim is to allow the airline companies to leverage the possibility to directly bargain online with customers in order to gather previously-inaccessible granular information. In doing so, they basically bypass the content aggregators scaling back the dominant role they have had. More precisely, as already experienced by OTAs, airline companies may benefit from the adoption of tracking mechanisms. As explained in [36
], online sellers may collect data in infinitely many ways: there are session-only tracking mechanisms, but also storage-based, cache-based, supercookies, fingerprinting. Browser cookies allow a web server to store a small amount of data on the devices of visiting users, which is then sent back to the web server upon subsequent request. Moreover, in line with [37
In conclusion, it is worth noting that the implementation of the NDC is not easy at it may seem: while airlines are relying on IT firms to build the proper original NDC, the GDSs are adopting the new standard too, and they are still maintaining their strong position (for a more detailed exposition, see for example [38
] or [39